Safety assurance tests reflecting actual accidents or near-misses are necessary to ensure the safety of automated driving systems. In such tests in a simulated environment, vehicle trajectories and road environments reproduced from nearmiss data recorded on a drive recorder are used. In this study, a method is proposed to reconstruct the trajectory of the ego vehicle and the road environment at the time the vehicle traveled from the dashcam video recorded on the drive recorder. The movement of the ego vehicle is obtained using image registration utilizing weights calculated from the semantic information obtained by semantic segmentation. Based on the ego vehicle's motion between frames calculated by image registration, the trajectory of the ego vehicle and the map of the driving environment are reconstructed. The validation of the proposed method on several dashcam videos suggests that it is possible to reconstruct the driving environment by image registration using semantic information.

View full abstract

In response to the poor performance of Visual SLAM in low light scenes and its high requirements for real-time computing, we propose a 2D-3D fused neural network that processes time consistency with minimal computational cost for enhancing a sequence of low-light images, and design two new loss functions to guide the network to focus on enhancing corners and edges. We then integrated our method into VINS-Mono and compared it with several low-light enhancement methods. The results show that our method achieves approximately 30% higher positioning accuracy than previous researches, while also having the shortest inference time.

View full abstract